For most of modern commerce, identity verification has been built around human interaction. A lawyer met a client, a bank checked a passport, a lender matched a name to a credit file and a hotel took a card at reception. Digital onboarding changed the channel, but not the assumption: behind the screen was a person asking another person or institution to do something. However, that assumption is starting to weaken - AI agents are moving from producing answers to carrying out tasks. The important change is not that software can write a convincing email. It is that software can search, compare, communicate, make choices and increasingly initiate transactions for a human or company. The legal challenge therefore moves beyond, “Is this person genuine?”, to “Which legal person is this machine acting for, what has it been authorised to do and who carries the loss if it exceeds that authority?”
The progression already visible
In August 2026, Google expanded AI Mode so users can describe a journey, compare current prices from more than 300 airlines and travel sites, track a chosen flight for price changes and build an itinerary. For hotels, users can describe their preferences, compare options and ask AI Mode to help complete the booking through integrated partners and Google Pay. The hotel or booking platform still handles the underlying booking and customer service, so this is not yet a machine independently signing every contract. But the direction is clear: AI is moving from search towards execution. The question is what happens when the same behaviour moves up the value chain? Asking software to find a £300 flight is low risk: asking it to find a £30,000 car, negotiate the price and arrange finance is different. And asking it to identify a £1 million house, negotiate an offer, instruct a conveyancer and arrange the movement of funds is different again. The technology merely has to become useful enough that humans delegate more decisions. Payments are already preparing for agents; the payment industry is not waiting for that future. Visa stated in July 2026 that AI agents were completing live purchases with participating European merchants, browsing products, selecting items and initiating purchases within cardholder-defined parameters. Mastercard has separately launched Agent Pay for Machines - designed for programmatic, always-on payments between software systems. In New Zealand, Mastercard and Westpac have demonstrated authenticated agentic transactions including a cinema-ticket purchase and an accommodation booking. These examples matter because payment changes the nature of the risk: a recommendation can be ignored, a payment creates an economic consequence. Once an agent can spend, the important evidence becomes the identity of the principal, the agent’s mandate, any spending limit, the goods or services it may buy and whether human approval is required above a threshold. That “the AI did it” is unlikely to be a satisfactory answer when money has moved to the wrong counterparty.
Meanwhile, real estate provides an even more useful case study. HomeLight launched EVA (an AI-powered escrow agent in the US) to automate parts of residential property closing. A typical escrow can involve more than 120 tasks: EVA can open orders, extract information from documents, obtain HOA records, interact with lenders, initiate title-related processes and use HomeLight’s corporate card to purchase documents. HomeLight asserts it has automated about a quarter of the escrow process, with difficult cases escalated rather than guessed. Important to note is that EVA is not an autonomous buyer deciding to purchase a house (and that distinction is important). Yet it demonstrates that AI agents are already entering workflows involving property, lenders, documents and money. If an agent can help process a property transaction, how long before a buyer instructs another agent to find properties within a mandate, compare surveys and financing, recommend an offer and begin the legal process? The final commitment may still require a human signature, but much of the journey to that point could be machine-led.
Furthermore, this is where the legal profession becomes more important, not less; the assumption that a convincing email, telephone voice or WhatsApp message identifies the client is already unsafe. In April 2026, the Solicitors Regulation Authority warned that AI had been used to manipulate or create identity documents and produce deepfakes designed to deceive liveness tests. Most attempts had failed, it said, but some had bypassed controls and, hence, the threat is current rather than theoretical. Suppose a long-standing corporate client’s authenticated AI agent emails its solicitor at 2am and asks for completion documents to be amended and £10 million released following completion. The identity may be genuine, but the harder questions are whether the company authorised the agent to alter the agreement, whether its mandate permits a £10 million instruction, whether that mandate has been revoked, whether board approval is required and whether the solicitor is entitled to rely on the machine’s instruction.
Identity, authority and liability are separate. A useful legal architecture is therefore: identify the principal, verify the agent’s mandate, validate that the proposed act falls within it record any required human approval, and then permit execution. Lawyers have spent centuries dealing with agency, powers of attorney, directors’ authority, trustees, fiduciaries and delegated powers - agentic AI does not abolish those concepts, it gives them a new digital form. Hence, law firms may consequently need a new kind of engagement record. A traditional retainer states in prose who the client is and what the firm will do, plus an agent-facing mandate may also need those fields a machine can test, such as:
· the legal principal
· the authorised agent
· the matter
· permitted actions
· maximum financial commitment
· whether payment authority exists
· actions reserved for a human expiry
· revocation status
That does not replace legal judgement, it renders authority easier to prove before a machine acts. Yet, it also raises questions about privilege. In April 2026, Sir Colin Birss, Chancellor of the High Court, used a speech on legal professional privilege in the age of AI to examine the distinction between legal advice from a lawyer and interactions with AI, referring to the Supreme Court’s Prudential decision and the continuing importance of confidentiality. For firms, the practical questions will increasingly be: if a client’s AI talks to a law firm’s AI, at what point is the communication part of the solicitor-client relationship, and what evidence proves the purpose, participants and authority behind it?
One way of thinking about the change is to separate three acts that human commerce often blurred together:
1. verified - who is behind the agent, and is that identity current?
2. validated - is the proposed transaction within the agent’s mandate and legally capable of proceeding?
3. vested - only then does money, property or another legal right move.
A passport check is not authority to buy a house, a recognised device is not authority to sign a guarantee and a valid payment credential is not proof that the payer intended this particular transaction. Hence, this separation could create a significant new role for lawyers: that is, rather than merely checking documents after machines have negotiated, lawyers could help define the boundaries within which agents may negotiate in the first place, such as transaction limits, jurisdictions, counterparties, reserved decisions, escalation rules and evidence standards. The legal product becomes partly preventative; machine-readable rules designed to stop an unauthorised transaction before it becomes litigation.
Notwithstanding, the economic consequences could be substantial. Agentic commerce can reduce search costs because software can compare far more suppliers than a person has time to examine. It can operate continuously, monitor prices and act when conditions are met. This could intensify competition, make pricing more transparent and allow smaller businesses to compete for customers they would never reach through conventional advertising. At the same time, it may weaken the value of familiar brands and paid search if machines increasingly choose suppliers using price, availability, verified credentials, insurance and contractual rules rather than prominence on a screen. Higher-value agentic transactions could also change professional services: estate agents, car dealers, mortgage brokers, travel agents, procurement teams and lawyers could spend less time collecting information and co-ordinating routine steps. Their value may move towards judgement, assurance, negotiation, exceptional cases and accepting responsibility. The productivity gain could be significant, but so could the redistribution of fees and employment. Firms whose economics depend on information friction may face pressure; firms that can make their products, authority and risk protections legible to machines may gain distribution.
The most important implication is therefore not that AI will replace lawyers - it is that lawyers may increasingly be asked to define what AI agents are allowed to do. A client may want an agent to negotiate but not sign, search but not purchase, spend £500 without approval but require a director above £50,000, or buy ordinary goods but never transfer land or grant security without a solicitor. Those are questions of authority, evidence and liability expressed as operational rules. Today, Google can monitor a flight price and help complete a hotel booking, Visa and Mastercard are building payment infrastructure for agents, HomeLight is putting AI agents inside residential property transactions, but none of those examples proves that autonomous agents will soon buy houses or companies without human involvement. However, they do show that the components are arriving separately: discovery, comparison, identity, payment and transaction workflow. The question for the legal profession is what happens when they join up: the next client may still be a human or company, but the instruction could arrive through software that looks, sounds and acts increasingly independently. Knowing the client will remain essential but it will simply no longer be enough. In the agentic economy, the valuable legal questions may become: who authorised the machine, what exactly was it allowed to do and can everyone prove the answer after the money has moved?
This article first appeared in Digital Bytes (27th of October, 2026), a weekly newsletter by Jonny Fry of Team Blockchain.
Who is responsible when an AI agent acts beyond the authority it was given?